Beyond the Data Cloud: Why Context Is the Real Bottleneck

"The era of passive observation is over."

That's how Google framed its Agentic Data Cloud announcement this month. It's the right framing. After two years of generative-AI demos that mostly produced summaries, search boxes, and the occasional hallucinated citation, the industry is finally turning toward agents that do things. Vodafone reports hundreds of agents running in production, saving "millions of euros every year." Virgin Voyages cut a mass itinerary rebook from six hours to eleven minutes.

The shift, in Google's own language, is from Systems of Intelligence (gen AI that answers) to Systems of Action (agents that perceive, reason, and execute). Gartner is calling 2026 "the year of agentic AI". We agree that's the inflection point. But agree-and-amplify is not the only honest response. The era of unguided action would be no improvement at all.

What Google got right

The architectural moves in this announcement are sound. The Knowledge Catalog (formerly Dataplex) is a serious attempt to give agents trustworthy, unified business context across data silos. The cross-cloud lakehouse with Apache Iceberg federation, Spanner Omni running outside Google Cloud, and Model Context Protocol support across BigQuery, AlloyDB, and Looker tell the same story: agents are useless when the data they need is locked in a different vendor's basement.

This is the right horizontal infrastructure. If your problem is moving petabytes of operational data across clouds without ETL pipelines and without paying egress tolls, the announcement matters to you. If your problem is teaching a Data Engineering Agent to write its own pipelines, the announcement matters to you.

But most enterprises do not actually have those problems.

The gap that remains

Even with unified data and capable agents, the Knowledge Catalog gives agents schema. It does not give them judgment.

A retail CFO does not need an agent that can join fourteen tables. She needs one that knows which signals matter this week, what the regulatory environment in Kenya looks like next month, and what the market narrative around her brand is doing in real time. None of that is a data problem. It is a context problem.

The same is true at smaller scales. A communications team launching a campaign does not need a faster query engine. It needs an agent that understands the difference between a journalist who will quote you accurately and one who will frame you as a punchline. That understanding is not in the data warehouse. It is in past coverage, current signals, and tacit institutional memory that no schema captures.

This is the layer above the data cloud. It is what we call contextual intelligence, and it is the part that most hyperscaler platforms quietly assume is somebody else's problem.

What the contextual layer looks like

We build it in three connected layers:

Internal context. The institutional memory of an organisation: documents, workflows, communications, regulatory filings, prior decisions. Most of this is unstructured and distributed across email, drives, and whatever was once called a wiki. Our ContentHub product turns it into a living context graph that agents can actually reason over. This is the same pattern retrieval-augmented generation (RAG) was supposed to solve, but with proper governance, lineage, and continuous enrichment baked in.

External signals. What the world is doing in real time. Media coverage, social conversations, regulatory shifts, market intent. PulseMI captures these continuously, with sentiment and intent layered on top, so that agents can act on what is happening rather than what was true last quarter. The closest analogue in the hyperscaler stack is real-time analytics, but the framing is different: signals over metrics.

Cross-domain synthesis. The interesting part. When you put internal context and external signals into the same reasoning system, patterns surface that neither could produce alone. The fact that a regulator quoted a specific phrase last Tuesday means one thing in isolation. It means something else when you know your competitor's general counsel posted a thread referencing the same phrase three days earlier. Agents that can reason across both surfaces start producing decisions that look less like search results and more like judgment.

This is what we mean by contextual intelligence: the layer where pattern recognition stops being statistical and starts being strategic.

Why this matters outside the Bay Area defaults

Hyperscaler platforms ship horizontal capability. They are not optimised for any specific market, regulatory environment, or industry. That is exactly what makes them useful as infrastructure, and exactly what makes them insufficient as intelligence.

A foundation model that has read the entire English-language internet still does not know which Kenyan parliamentary committee is currently reviewing a finance bill, what the political alignment of its chair is, or how the Daily Nation framed the last three iterations of similar legislation. A general-purpose agent built on that foundation will guess. A contextually grounded agent will know.

This is not just an African insight. It applies anywhere a market has its own regulatory context, language nuances, distribution realities, or institutional history. Which is to say: everywhere outside a small handful of commodity sectors. The Stanford AI Index 2025 flagged the same pattern: foundation models systematically under-represent non-Anglophone, non-Western contexts, and downstream errors compound at scale.

Our work in Kenya, where local regulation, local media, and local market structure differ meaningfully from any default model's training distribution, is a useful test case for the broader thesis. The contextual layer becomes a defensible advantage precisely because it is not in the model. The Kenya Data Protection Act, 2019 is a single example of regulatory context that no Bay Area model carries by default but every Kenyan enterprise must operate within.

Two layers, both necessary

Google announced the data infrastructure. We are building the layer above it.

Both will exist. Both are necessary. But they answer different questions. Their question is how do we move information without friction across clouds and into agent memory? Ours is how do we make sure the agent's memory contains the right things, with the right weight, for the decision in front of it?

If you are running a data platform team, the announcement is an upgrade. If you are running a business that needs to make decisions with conviction, the announcement is necessary but not sufficient. The intelligence still has to come from somewhere.

That somewhere is the contextual layer. And that is where we are building.

---

Want to see what contextual intelligence looks like in practice? Talk to us, or explore the platform.

Further reading

• Google Cloud, What's new in the Agentic Data Cloud — the announcement this commentary responds to.
• Anthropic, Introducing the Model Context Protocol — the open standard for connecting agents to data sources, now adopted across major platforms.
• McKinsey, The state of AI in early 2025 — survey data on enterprise AI adoption beyond the headlines.
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — the foundational paper on RAG, which this contextual-layer thinking extends.
• Apache Iceberg, Project documentation — the open table format underpinning the cross-cloud lakehouse pattern.
• Stanford HAI, The 2025 AI Index Report — annual measurement of AI capability, deployment, and limitations.
• Kenya Office of the Data Protection Commissioner, Data Protection Act compliance — referenced for the regional regulatory context.

Related from Context Engine

• ContentHub: Document Intelligence — the internal-context layer in production.
• PulseMI: Real-time Media Intelligence — the external-signals layer for African markets.
• InContext PR — the synthesis layer applied to PR and communications workflows.
• About Context Engine — why we are building the contextual layer above the data cloud.